feat(s2): elementwise unary-math kernels (exp/log/sqrt/erf/…) on gfx1151 + AVX-512

docs/audit/backend/rocm/ROCM_AUDIT.md

@@ -341,6 +341,21 @@ become the on-silicon **oracle** the compiled path validates against.
   GPU-free codegen gate. The CPU half (AVX-512) landed in the x86 backend
   (`avx512_reduce_f32.cpp`) — so the reduction family now has a real optimized
   kernel on both devices we have hardware for. Status `compiled`.
+- **Elementwise unary math** (S2 scalar-math / stability family, 2026-06-25):
+  a `tessera_rocm.unary` directive + `generate-rocm-unary-kernel` pass emitting
+  a flat per-element kernel (one thread per element), the unary sibling of the
+  activation lane. Covers `exp`/`log`/`sqrt`/`rsqrt`/`reciprocal`/`abs`
+  (`absolute`)/`sign`/`erf`/`tanh`/`sigmoid`/`log1p`/`expm1`/`softplus`
+  (softplus stable: `log1p(exp(-|x|)) + max(x,0)`); transcendentals lower
+  through the `math` → ROCDL path. New `runtime.launch()` lane
+  `rocm_unary_compiled`, dispatched by op name; f16/bf16/f32 storage, f32
+  compute. Validated on gfx1151 vs numpy across kind × dtype × shape incl.
+  rank-3 (`test_rocm_unary_compiled.py`) + a GPU-free codegen gate. The CPU
+  half landed in the x86 backend as an AVX-512 kernel for the **algebraic
+  subset** (`sqrt`/`rsqrt`/`reciprocal`/`abs`/`neg`/`sign`, direct intrinsics,
+  no polynomial approx — `avx512_unary_f32.cpp`, validated standalone); the
+  transcendentals stay numpy-reference on CPU (no fused x86 claim). Status
+  `compiled`.
 - **rmsnorm / layer_norm** (2026-06-25): the row-reduction siblings of the
   softmax kernel — a `tessera_rocm.norm` directive + `generate-rocm-norm-kernel`
   pass (one workgroup per row). rmsnorm tree-reduces Σx² in one pass; layer_norm

docs/audit/generated/runtime_abi.csv

@@ -578,8 +578,10 @@ x86,tessera_x86_amx_gemm_s8s8_s32,amx_gemm_s8s8_s32,,src/compiler/codegen/tesser
 x86,tessera_x86_amx_gemm_s8s8_s32,amx_gemm_s8s8_s32,,src/compiler/codegen/tessera_x86_backend/src/kernels/amx_gemm_int8.cpp
 x86,tessera_x86_avx512_gemm_bf16,avx512_gemm,bf16,src/compiler/codegen/tessera_x86_backend/src/kernels/avx512_gemm_bf16.cpp
 x86,tessera_x86_avx512_reduce_f32,avx512_reduce,f32,src/compiler/codegen/tessera_x86_backend/src/kernels/avx512_reduce_f32.cpp
+x86,tessera_x86_avx512_unary_f32,avx512_unary,f32,src/compiler/codegen/tessera_x86_backend/src/kernels/avx512_unary_f32.cpp
 x86,tessera_x86_avx512_vnni_gemm_u8s8_s32,avx512_vnni_gemm_u8s8_s32,,src/compiler/codegen/tessera_x86_backend/src/kernels/avx512_vnni_gemm_int8.cpp
 x86,tessera_x86_epilogue_bias_fp32,epilogue_bias_fp32,,src/compiler/codegen/tessera_x86_backend/src/kernels/epilogue.cpp
 x86,tessera_x86_epilogue_bias_gelu_fp32,epilogue_bias_gelu_fp32,,src/compiler/codegen/tessera_x86_backend/src/kernels/epilogue.cpp
 x86,tessera_x86_reference_gemm_bf16,reference_gemm,bf16,src/compiler/codegen/tessera_x86_backend/src/kernels/avx512_gemm_bf16.cpp
 x86,tessera_x86_reference_reduce_f32,reference_reduce,f32,src/compiler/codegen/tessera_x86_backend/src/kernels/avx512_reduce_f32.cpp
+x86,tessera_x86_reference_unary_f32,reference_unary,f32,src/compiler/codegen/tessera_x86_backend/src/kernels/avx512_unary_f32.cpp

docs/audit/generated/runtime_abi.md

@@ -4,7 +4,7 @@ Human-readable view. The canonical machine-readable artifact is `runtime_abi.csv
 
 ## Headline
 
-- **328** unique `extern "C" tessera_*` C ABI symbols across all backends.
+- **330** unique `extern "C" tessera_*` C ABI symbols across all backends.
 - **6 / 6** core runtime headers present.
 - **134** Apple GPU kernel families with per-dtype variants.
 
@@ -26,7 +26,7 @@ Human-readable view. The canonical machine-readable artifact is `runtime_abi.csv
 | `apple` | 304 |
 | `nvidia` | 4 |
 | `rocm` | 10 |
-| `x86` | 10 |
+| `x86` | 12 |
 
 ## Apple GPU kernel families × dtype matrix
 

docs/audit/generated/runtime_execution_matrix.csv

@@ -19,4 +19,5 @@ rocm,rocm_reduce_compiled,native_gpu,1,rocm_reduce_compiled,success,hip_runtime,
 rocm,rocm_rope_compiled,native_gpu,1,rocm_rope_compiled,success,hip_runtime,"ROCm rope artifact runs the COMPILER-GENERATED interleaved-pair rotary-position-embedding kernel (one workgroup per row): tessera-opt generates + serializes the kernel to hsaco in-process, then HIP loads + launches it."
 rocm,rocm_silu_mul_compiled,native_gpu,1,rocm_silu_mul_compiled,success,hip_runtime,"ROCm silu_mul artifact runs the COMPILER-GENERATED flat 2-operand elementwise SwiGLU gate-multiply silu(a)·b (one thread per element): tessera-opt generates + serializes the kernel to hsaco in-process, then HIP loads + launches it."
 rocm,rocm_softmax_compiled,native_gpu,1,rocm_softmax_compiled,success,hip_runtime,"ROCm softmax artifact runs the COMPILER-GENERATED RDNA row-reduction kernel (stable softmax over the last axis, one workgroup per row, LDS tree-reduce): tessera-opt generates + serializes the kernel to hsaco in-process, then HIP loads + launches it. The first non-matmul/non-WMMA compiled ROCm kernel."
+rocm,rocm_unary_compiled,native_gpu,1,rocm_unary_compiled,success,hip_runtime,"ROCm unary artifact runs the COMPILER-GENERATED flat elementwise unary-math kernel (S2 scalar-math/stability: exp/log/sqrt/rsqrt/reciprocal/abs/sign/erf/tanh/sigmoid/log1p/expm1/softplus, one thread per element): tessera-opt generates + serializes the kernel to hsaco in-process, then HIP loads + launches it. Dispatched by op name."
 rocm,rocm_wmma,native_gpu,1,rocm_wmma,success,hip_runtime,"ROCm matmul via the hand-written RDNA WMMA GEMM (tessera_rocm_wmma_gemm_{f16,bf16} C ABI symbol, HIPRTC-compiled for the device arch). Now the reference ORACLE + availability fallback for the compiled lane (rocm_compiled) — still directly selectable by stamping compiler_path=""rocm_wmma""."

docs/audit/generated/runtime_execution_matrix.md

@@ -33,6 +33,7 @@ Single source of truth for what `runtime.launch()` does with each `(target, comp
 | `rocm` | `rocm_rope_compiled` | `rocm_rope_compiled` | `native_gpu` | `hip_runtime` | ROCm rope artifact runs the COMPILER-GENERATED interleaved-pair rotary-position-embedding kernel (one workgroup per row): tessera-opt generates + serializes the kernel to hsaco in-process, then HIP loads + launches it. |
 | `rocm` | `rocm_silu_mul_compiled` | `rocm_silu_mul_compiled` | `native_gpu` | `hip_runtime` | ROCm silu_mul artifact runs the COMPILER-GENERATED flat 2-operand elementwise SwiGLU gate-multiply silu(a)·b (one thread per element): tessera-opt generates + serializes the kernel to hsaco in-process, then HIP loads + launches it. |
 | `rocm` | `rocm_softmax_compiled` | `rocm_softmax_compiled` | `native_gpu` | `hip_runtime` | ROCm softmax artifact runs the COMPILER-GENERATED RDNA row-reduction kernel (stable softmax over the last axis, one workgroup per row, LDS tree-reduce): tessera-opt generates + serializes the kernel to hsaco in-process, then HIP loads + launches it. The first non-matmul/non-WMMA compiled ROCm kernel. |
+| `rocm` | `rocm_unary_compiled` | `rocm_unary_compiled` | `native_gpu` | `hip_runtime` | ROCm unary artifact runs the COMPILER-GENERATED flat elementwise unary-math kernel (S2 scalar-math/stability: exp/log/sqrt/rsqrt/reciprocal/abs/sign/erf/tanh/sigmoid/log1p/expm1/softplus, one thread per element): tessera-opt generates + serializes the kernel to hsaco in-process, then HIP loads + launches it. Dispatched by op name. |
 | `rocm` | `rocm_wmma` | `rocm_wmma` | `native_gpu` | `hip_runtime` | ROCm matmul via the hand-written RDNA WMMA GEMM (tessera_rocm_wmma_gemm_{f16,bf16} C ABI symbol, HIPRTC-compiled for the device arch). Now the reference ORACLE + availability fallback for the compiled lane (rocm_compiled) — still directly selectable by stamping compiler_path="rocm_wmma". |
 
 ## Targets with no executable row
@@ -67,4 +68,5 @@ nvidia_sm80, nvidia_sm90, nvidia_sm100, rocm_gfx90a, rocm_gfx940, rocm_gfx942, r
 | `rocm_rope_compiled` | AMD GPU RDNA rotary-position-embedding the Tessera compiler GENERATES (generate-rocm-rope-kernel -> ROCDL -> hsaco, in-process via tessera-opt), then HIP loads + launches it. Interleaved-pair RoPE over [M, D] (one workgroup per row); f32/f16/bf16 |
 | `rocm_silu_mul_compiled` | AMD GPU RDNA SwiGLU gate-multiply the Tessera compiler GENERATES (generate-rocm-silu-mul-kernel -> ROCDL -> hsaco, in-process via tessera-opt), then HIP loads + launches it. Flat 2-operand elementwise silu(a)·b (one thread per element); the standalone analog of the fused SwiGLU gate-multiply; f32/f16/bf16 storage, f32 compute |
 | `rocm_softmax_compiled` | AMD GPU RDNA row-reduction softmax the Tessera compiler GENERATES (generate-rocm-softmax-kernel -> ROCDL -> hsaco, in-process via tessera-opt), then HIP loads + launches it. Stable softmax over the last axis (one workgroup per row, LDS tree-reduce); the first non-matmul/non-WMMA compiled ROCm kernel. f32/f16/bf16 storage, f32 reduce |
+| `rocm_unary_compiled` | AMD GPU RDNA flat elementwise unary-math kernel the Tessera compiler GENERATES (generate-rocm-unary-kernel -> ROCDL -> hsaco, in-process via tessera-opt), then HIP loads + launches it — the S2 scalar-math / stability family (exp/log/sqrt/rsqrt/reciprocal/abs/sign/erf/tanh/sigmoid/log1p/expm1/softplus), one thread per element, dispatched by op name; f32/f16/bf16 storage, f32 compute |
 | `rocm_wmma` | AMD GPU RDNA WMMA matrix-core GEMM via the shipped libtessera_rocm_gemm.so tessera_rocm_wmma_gemm_{f16,bf16} C ABI symbol (HIPRTC-compiled for the device arch; f16/bf16 storage, f32 accumulate) |

docs/audit/generated/test_coverage.csv

@@ -1,6 +1,6 @@
 op,python_refs,lit_refs,negative_refs,total_refs,is_thinly_tested,dtype_variants,bucket,reason
-abs,2,0,0,2,0,,directly_tested,2 direct test references
-absolute,1,0,0,1,1,,structural_only,unclassified — defaults to structural_only
+abs,3,0,0,3,0,bf16 f32,directly_tested,3 direct test references
+absolute,2,0,0,2,0,bf16 f32,directly_tested,2 direct test references
 acos,1,0,0,1,1,,covered_by_family,category default for 'elementwise'
 adafactor,4,0,0,4,0,fp32 fp64,directly_tested,4 direct test references
 adam,7,5,0,12,0,fp32 fp64,directly_tested,12 direct test references
@@ -171,11 +171,11 @@ ema_update,1,0,0,1,1,,structural_only,category default for 'grad_transform'
 empty_state_tree,0,0,0,0,1,,structural_only,category default for 'state_tree'
 eq,0,0,0,0,1,,covered_by_family,category default for 'comparison'
 equiprob_band_partition,0,0,0,0,1,,structural_only,unclassified — defaults to structural_only
-erf,1,0,0,1,1,,covered_by_family,category default for 'elementwise'
+erf,2,0,0,2,0,bf16 f32,directly_tested,2 direct test references
 erfc,1,0,0,1,1,,covered_by_family,category default for 'elementwise'
-exp,4,0,0,4,0,,directly_tested,4 direct test references
+exp,5,0,0,5,0,bf16 f32,directly_tested,5 direct test references
 expand,1,0,0,1,1,,structural_only,unclassified — defaults to structural_only
-expm1,1,0,0,1,1,,covered_by_family,category default for 'elementwise'
+expm1,2,0,0,2,0,bf16 f32,directly_tested,2 direct test references
 factorized_matmul,4,0,0,4,0,,directly_tested,4 direct test references
 factorized_pos_emb,6,0,1,6,0,,directly_tested,6 direct test references
 fake_quantize,3,0,0,3,0,,directly_tested,3 direct test references
@@ -247,8 +247,8 @@ lion,4,0,0,4,0,fp32 fp64,directly_tested,4 direct test references
 load_balance_loss,2,0,0,2,0,,directly_tested,2 direct test references
 load_sharded,3,0,1,3,0,,directly_tested,3 direct test references
 load_state,14,0,2,14,0,,directly_tested,14 direct test references
-log,3,0,0,3,0,,directly_tested,3 direct test references
-log1p,2,0,0,2,0,,directly_tested,2 direct test references
+log,4,0,0,4,0,bf16 f32,directly_tested,4 direct test references
+log1p,3,0,0,3,0,bf16 f32,directly_tested,3 direct test references
 log_cosh_loss,1,0,0,1,1,,covered_by_family,category default for 'loss'
 log_softmax,5,5,0,10,0,,directly_tested,10 direct test references
 logical_and,1,0,0,1,1,,structural_only,unclassified — defaults to structural_only
@@ -345,7 +345,7 @@ quantize_int8,3,0,0,3,0,,directly_tested,3 direct test references
 quantize_nvfp4,11,0,2,11,0,fp16 fp32 fp4_e2m1 fp6_e2m3 fp8_e4m3 fp8_e5m2 int8 nvfp4,directly_tested,11 direct test references
 quantized_matmul,5,0,0,5,0,f16,directly_tested,5 direct test references
 rearrange,2,0,0,2,0,,directly_tested,2 direct test references
-reciprocal,1,0,0,1,1,,structural_only,unclassified — defaults to structural_only
+reciprocal,2,0,0,2,0,bf16 f32,directly_tested,2 direct test references
 reduce,86,0,7,86,0,f32 fp16 fp32 fp4_e2m1 fp6_e2m3 fp8_e4m3 nvfp4,directly_tested,86 direct test references
 reduce_scatter,0,4,0,4,0,,directly_tested,4 direct test references
 relu,105,20,9,125,0,bf16 f16 f32 f64 fp32 fp8_e4m3 int8,directly_tested,125 direct test references
@@ -381,7 +381,7 @@ rope,19,10,0,29,0,bf16,directly_tested,29 direct test references
 rope_merge,2,0,1,2,0,fp16 fp32,directly_tested,2 direct test references
 rope_split,8,0,1,8,0,fp16 fp32,directly_tested,8 direct test references
 round,1,0,0,1,1,,structural_only,unclassified — defaults to structural_only
-rsqrt,2,0,0,2,0,,directly_tested,2 direct test references
+rsqrt,3,0,0,3,0,bf16 f32,directly_tested,3 direct test references
 safetensors_export,1,0,0,1,1,,structural_only,category default for 'aot'
 save_sharded,1,0,1,1,1,,structural_only,category default for 'serialization'
 save_state,6,0,2,6,0,,directly_tested,6 direct test references
@@ -399,9 +399,9 @@ seq2seq_loss,3,0,0,3,0,,directly_tested,3 direct test references
 sgd,6,0,0,6,0,,directly_tested,6 direct test references
 shard_map,0,0,0,0,1,,structural_only,category default for 'sharding'
 sharded_dataset,2,0,0,2,0,,directly_tested,2 direct test references
-sigmoid,4,0,0,4,0,bf16 f16 f32 fp32,directly_tested,4 direct test references
+sigmoid,5,0,0,5,0,bf16 f16 f32 fp32,directly_tested,5 direct test references
 sigmoid_safe,4,0,0,4,0,,directly_tested,4 direct test references
-sign,2,0,0,2,0,,directly_tested,2 direct test references
+sign,4,0,1,4,0,bf16 f32,directly_tested,4 direct test references
 silu,84,2,5,86,0,bf16 f16 f32 f64 fp16 fp32 fp4_e2m1 fp6_e2m3 fp8_e4m3 nvfp4,directly_tested,86 direct test references
 silu_mul,14,13,0,27,0,bf16 fp32,directly_tested,27 direct test references
 simple_rnn_cell,5,0,0,5,0,,directly_tested,5 direct test references
@@ -410,17 +410,17 @@ sinh,1,0,0,1,1,,covered_by_family,category default for 'elementwise'
 slice,11,0,0,11,0,,directly_tested,11 direct test references
 smooth_l1_loss,1,0,0,1,1,,covered_by_family,category default for 'loss'
 softcap,7,0,1,7,0,,directly_tested,7 direct test references
-softmax,97,37,9,134,0,bf16 f16 f32 fp16 fp32 fp4_e2m1 fp6_e2m3 fp8_e4m3 nvfp4,directly_tested,134 direct test references
+softmax,98,37,10,135,0,bf16 f16 f32 fp16 fp32 fp4_e2m1 fp6_e2m3 fp8_e4m3 nvfp4,directly_tested,135 direct test references
 softmax_safe,4,4,1,8,0,bf16,directly_tested,8 direct test references
-softplus,2,0,0,2,0,,directly_tested,2 direct test references
+softplus,3,0,0,3,0,bf16 f32,directly_tested,3 direct test references
 sort,3,0,0,3,0,,directly_tested,3 direct test references
 spectral_conv,5,0,0,5,0,fp32,directly_tested,5 direct test references
 spectral_filter,3,0,0,3,0,fp32,directly_tested,3 direct test references
 spectral_norm,2,0,0,2,0,,directly_tested,2 direct test references
 split,1,0,0,1,1,,structural_only,unclassified — defaults to structural_only
 spmm_coo,3,0,0,3,0,,directly_tested,3 direct test references
 spmm_csr,2,0,0,2,0,,directly_tested,2 direct test references
-sqrt,4,0,0,4,0,,directly_tested,4 direct test references
+sqrt,5,0,0,5,0,bf16 f32,directly_tested,5 direct test references
 squeeze,1,0,0,1,1,,structural_only,unclassified — defaults to structural_only
 stablehlo_export,2,0,0,2,0,,directly_tested,2 direct test references
 stack,2,0,0,2,0,,directly_tested,2 direct test references
@@ -437,7 +437,7 @@ svd,8,5,0,13,0,bf16 f16 f32 fp16 fp32,directly_tested,13 direct test references
 switch,0,0,0,0,1,,structural_only,category default for 'control_flow'
 take,1,0,0,1,1,,structural_only,unclassified — defaults to structural_only
 tan,1,0,0,1,1,,covered_by_family,category default for 'elementwise'
-tanh,4,0,1,4,0,bf16 f16 f32,directly_tested,4 direct test references
+tanh,5,0,1,5,0,bf16 f16 f32,directly_tested,5 direct test references
 tile,1,0,0,1,1,,structural_only,unclassified — defaults to structural_only
 tile_view,2,0,0,2,0,,directly_tested,2 direct test references
 tiny_attention_conformance,0,0,0,0,1,,structural_only,category default for 'conformance'